MACE-MP-0: A Universal Machine-Learned Interatomic Potential

MACE-MP-0 is a foundation model for atomistic materials chemistry built on the MACE architecture, which unifies the atomic cluster expansion with equivariant message passing.

Architecture: High body-order (four-body) equivariant features in every layer mean only two message-passing layers are needed, making the model markedly cheaper to train and evaluate than earlier equivariant potentials while reaching state-of-the-art accuracy on benchmarks such as rMD17 and 3BPA.

Training data: Trained on MPtrj, a public dataset of roughly 150k inorganic crystal configurations from the Materials Project, covering most of the periodic table under the PBE functional.

Reported behaviour: Runs stable molecular dynamics out of the box for solids, liquids, gases, chemical reactions, interfaces and even a small protein, with qualitative and sometimes quantitative accuracy; fine-tuning on a small number of application-specific configurations recovers ab initio accuracy. Typical throughput is around a thousand atoms for nanoseconds per day on a single GPU.

Industrial relevance: It removes the per-system potential development cycle that has historically blocked ML potentials from production use. Code is released on GitHub under the MIT licence.